The recognition result is very bad,such as the following iamges

[Jayhello]

result is like that:
Recognized text: vaossyoba (raw: v--a----o-ss-s--y-o-b--a--)

and many simple image ,the result is very bad? anyone can tell me why?

[bgshih]

The model is trained on a dataset that is heavily biased towards a-z characters. Training on a balanced dataset would help.

[rayush7]

@bgshih If I want the model to recognize words and multi-digit numbers both then will fine-tuning the model (already trained on VGG Synthetic word dataset) on dataset like Street View House Number (SVHN) gonna help? Or I need to retrain the model from the scratch?

[bgshih]

@rayush7 I strongly feel fine-tuning be a better choice than starting from scratch.

[rayush7]

Thanks @bgshih

[rayush7]

@bgshih Could you please explain the format of the single/multi digit labels for creating lmdb datasets?
Example if there is a word "Apple" in the image then setting its groundtruth label as the string 'Apple' is working for me. I can convert it into lmdb format and create my train and validation set and train my own model.
But if I have numerical labels for example number 350 in the image then by setting its label to string '350' is giving me error in the ascii2label function (in utilities.lua). Even though in ascii2label function if conditions are present to take into account the numerical digit values but still its throwing the error. Please explain what should be the groundtruth format of the single/multi digit labels for creating lmdb datasets?

[bgshih]

@rayush7 This is not an expected behavior. What was the error message and where was it thrown?

[rayush7]

@bgshih Hey the problem was in my lmdb dataset and not in code. The label string was in unreadable form. I resolved this problem. Now giving the '350' as a label for a image with number 350 in it works like a charm! Thanks @bgshih

[rayush7]

@bgshih I am trying to finetune the crnn model (pretrained on vgg synthetic dataset) on svhn dataset. The images (32x32) I am using look like these

While finetuning, the training loss is showing random patterns and no legitimate inference can be drawn because of that. Also all the validation set images are classified as "1". I am unable to figure out if this is a problem due to wrong choice of optimization parameters or crnn is unfit for training on such images. I am using the default config.lua file with adadelta as the optimizer. Any suggestions on what mistake I am making and how can I rectify that?

[ahmedmazari-dhatim]

@bgshih , @rayush7 can you tell me the steps you followed to fine tune the pre-trained mode with your own dataset. l'm newbie on fine tuning pre-trained model. l want to try it for the first time .

Thank you

[alexiskattan]

@rayush7 Did you train the model with digits? If so, could you post it online?